Research on Infrared Image Simulation Method Based on Multi-auxiliary Task Style Transfer Networks

Unsupervised style transfer network can accomplish the task of infrared image simulation, and the core of its goal is to migrate the style of infrared images into visible images while maintaining the content of visible images. However, the existing methods have problems such as content loss, edge blurring, and poor stylization effect. In this paper, Multi-auxiliary Task Style Migration Network (MATST) is proposed. The problems of content loss and edge blurring are solved by adding coordinated attention to the generator of the generative adversarial network. The content simulation effect is further improved by adding semantic segmentation auxiliary task. Solve the problem of poor stylization effect by multi-layer style mapping auxiliary task. Enhance model robustness and generalization by domain enhancement auxiliary task. Finally, a comparison is made with other methods in terms of qualitative and quantitative indexes. It surpasses the original method in objective metrics, with 30% reduction in Style loss [1] and 5.6% reduction in LPIPS [2] . It is also more realistic in subjective visualization.